Building Sublett Labs: From NICU to AI Agents
I'm a respiratory therapist. I work 12-hour shifts in a NICU. That's my actual job — the one that pays the bills, the one with benefits, the one I can't quit because I have a family and a mortgage like everyone else.
Sublett Labs is what I do after shift. Nights, weekends, the random Tuesday at 2am when I can't sleep. It's not a side hustle in the "buy my course" sense. It's me trying to build things I need, that turn into things other people might need.
Here's how it actually started, what I built, and what I've learned doing both jobs at the same time.
The Problem Was Mine First
I run a small e-commerce store. Nothing huge — print-on-demand, dropshipping, the usual. But it generates enough that missing a call matters. A customer wants to know about an order. A supplier has a question. Someone calls about a wholesale opportunity.
The problem: I can't answer the phone when I'm in the NICU. You don't step away from a ventilator to take a sales call. And after hours, I'm either asleep or spending time with my family. Voicemail was catching everything. And nobody leaves voicemails anymore.
I looked at answering services. $200-400/month for a human who doesn't know my business, reads from a script, and forwards maybe 30% of calls to me. I looked at virtual receptionist platforms — cheaper, but robotic and limited. Neither felt right.
So I built one. That was the start.
Clara Was First
Clara is an AI receptionist. She answers calls, books appointments, answers questions about the business, and knows when to say "I don't know." I built her because I needed her.
The first version was terrible. Bad latency, robotic voice, hallucinated answers. I spent three months on the "I don't know" problem alone — teaching the model to recognize when a question is outside its scope and route the call instead of making something up. That's the part most AI answering services still haven't solved.
Clara works now. She handles my calls. That's the only customer I can claim with certainty — me. I'm not going to pretend there's a roster of paying clients. There isn't. I built it for myself, it works for me, and I'm now trying to see if it works for other people too.
Then the Other Agents
Once Clara was running, I noticed a pattern. Building one AI agent means building infrastructure — monitoring, logging, API management, a way to register and discover agents. That infrastructure doesn't come with the agent. You build it yourself.
So I built more agents, each solving a problem I had:
- ResumeForge — I tried AI resume builders and they all invented facts I didn't claim. So I built one that doesn't fabricate. If you didn't say it, it won't appear.
- AgentSeek — I had multiple agents running and couldn't keep track of their endpoints. Built a registry — DNS for AI agents.
- Agent Monitor — My agents kept dying at 3am and I'd find out hours later. Built a heartbeat monitor that alerts me when something goes down.
- Clozr — I was missing follow-ups because I couldn't keep track of where every lead was in the pipeline. Built a lightweight CRM that actually closes leads.
- AgentPay — I needed a way to manage API credits across agents without manual tracking. Built a spend tracker.
Six agents. All built because I had the problem. None of them have paying customers yet. That's the honest truth.
What I'd Tell Myself on Day One
If I could go back to when I started, here's what I'd say:
Build the boring infrastructure first. I spent months on Clara's voice and latency before I had monitoring in place. That meant I was debugging production issues blind. A simple heartbeat and logging setup would have saved me weeks of guessing.
Don't try to do everything locally. I wanted to run everything on my own hardware for cost reasons. Whisper on the GPU worked. The 70B LLM did not — 500ms+ to first token on a phone call is a non-starter. Groq's dedicated hardware was the only way to hit sub-1s latency. Sometimes the cloud is the right answer.
Stop building new agents and start selling one. This is the lesson I'm still learning. It's more fun to build than to sell. But six agents with zero revenue is just a hobby. I need to pick one — probably Clara — and figure out how to get it into other people's hands.
Write things down. I lost track of decisions, configurations, and lessons learned because I wasn't keeping notes. Memory files, build logs, daily notes — they've saved me more times than I can count.
The NICU Thing
People ask how I balance the two. Honestly, I don't always. There are weeks where I barely touch the computer because the hospital is slammed. There are shifts where I'm thinking about a bug in Clara while I'm adjusting a ventilator. Neither job gets 100% of me.
But here's the thing — the NICU made me a better builder. In respiratory therapy, if you make a mistake, a premature baby dies. That changes how you approach software. I test everything. I deploy carefully. I build in fallbacks. I expect things to fail because in my day job, failure isn't an option.
The urgency is different too. In the NICU, you act now and think later. In software, I can take my time. That's a luxury. But it also means I sometimes overthink things that should just ship. I'm working on that.
What's Next
Right now, Sublett Labs is me, a server in my house, two GPUs, and six agents that I built for myself. The next step is the hard part — getting them into other people's hands. Not through hype or cold email blasts, but by showing the work.
I'm writing this blog to document the build process. The technical decisions, the failures, the lessons. If you're building something similar — an AI agent, a SaaS, a side project while working a full-time job — maybe some of this is useful.
If not, that's fine too. I'm mostly writing it for myself anyway.
← Back to blog